Open source has come a long way over the last few years, and most of the tools I’ve tried are either level with their closed counterparts or genuinely better in some ways. Most of them are free, and the transparency never really wears off for me. Plus a lot of these projects end up with weird little features the paid tools would never bother shipping.
So most of my paid stack has been replaced over the past years or two and I don’t miss much of it. But there’s one category where I keep pulling my card back out no matter how many alternatives I test, and it’s a category I actually really wanted open source to solve, but it just isn’t there yet.
And none of it felt like a downgrade
Notion was the first to go, replaced with Affine, which is genuinely one of the most impressive open-source apps I’ve come across. It’s block-based like Notion but with an infinite whiteboard baked into the same canvas as your docs, so a page can be a document one minute and a diagram the next without having to switch to another app. Notion still doesn’t have this in 2026.
Then there’s Readwise, which I swapped for Wallabag. Wallabag is a read-it-later tool with highlights and tagging and a browser extension for saving as you go. You can also self-host it if you want. I do miss Readwise’s spaced-repetition review, but I’m not actively studying my highlights every day so it hasn’t been a real loss.
Acrobat also went out the window at some point. OmniTools replaced it and it actually does way more than Acrobat. All the merging and splitting and OCR stuff is there, but so is password protection and conversion between formats (for free), among many other non-PDF tools such as basic image and video editing. The whole suite of tools can be self hosted and it’s super lightweight to run.
I also cancelled my image and design subscriptions a long time ago now, and finding alternatives has been quite a breeze. Krita, Penpot, and Inkscape are just a few worth naming. Though I don’t use them all every single day, I keep them pinned in my task bar.
The category open-source hasn’t caught up on
It’s mostly about horsepower
The one category I can’t quit paying for is AI. And specifically, cloud AI chatbots and vibe design tools. That’s where my Claude Pro sub goes, and where Google AI Plus goes for Gemini and NotebookLM. I’ve tried to build an equivalent stack out of open-source pieces, and for some use cases they come close and are enough, but it’s just not enough for serious work or projects.
Cloud AI just has the horsepower that open-weights don’t. Claude Opus and Gemini Pro are running in data centers on hardware that costs more than most cars, with parameter counts my RTX 3070 can’t even load a fraction of. There’s just no consumer hardware that competes with that. My GPU caps out around 12B at reasonable quants, and even the best models at that size feel noticeably shallower on heavy projects.
Context length is the other one. Claude gives me 200k tokens out of the box and Gemini pushes past a million, and I can actually use them without watching for out-of-memory errors. Local models boast long contexts too, but the amount you can actually load into VRAM alongside the KV cache is way less than what the model advertises, depending on your hardware.
Then there’s everything the interfaces bring on top of the model itself. Projects and Artifacts, proper document parsing on file uploads, a deep research mode that actually digs through sources, and memory that carries across chats. On the open-source side, most of these are separate MCP servers or plugins or Docker containers, and stringing them together is its own weekend project. NotebookLM specifically does source-grounded synthesis with citations across dozens of documents, and Open Notebook exists but is a bit of a pain to set up (the closest open-source equivalent to NotebookLM I can recommend is SurfSense).
Claude Design is cloud-only too, and there’s nothing self-hosted that does what it does for design-to-code work. That’s my main design tool for anything vibe-based lately. Open Design and Open CoDesign come close, but only with cloud APIs; they lag severely when you hook up a local model.
Overall, I just get more reliability with cloud AI; there’s no debugging or extra setup required. I don’t even use my open-source productivity stack as much lately anyway, not since setting up Claude Code to organize my notes and manage my files for me.
It wasn’t for a lack of trying
Local models have their place for productivity
It’s not for a lack of trying, though. I genuinely love running local LLMs. There’s something novel about downloading weights and having a real conversation with something running entirely on your own machine. It’s a hobby as much as a tool for me at this point.
And they do have real uses. Anything private goes to local by default, whether that’s financial stuff or medical questions or personal drafts I don’t want to send out to a server. They’re also a solid backup when my internet drops or I’m working somewhere without a reliable connection.
But the ceiling is hard. On my little VRAM, I’m never going to touch Opus or even Sonnet level reasoning, no matter which small model I’m running or how much I fine-tune my prompt. As mentioned, context length in practice is also much smaller than what the model card advertises. And every basic feature you’d expect from a chatbot interface becomes a mini setup project of its own – want web search? Set up Brave Search MCP. Want vision? Hunt for a whole new model. Want document parsing? Add a whole other tool.
So local LLMs sit in my stack as a complement, not a replacement. They handle the private and offline use cases very well, and I use them nearly every day for regular tasks too. But Cloud is where I’m actually going to get the depth and speed I need.
Almost fully open source
Open source really isn’t the downgrade it used to be. Look at Affine, or Penpot on the design side – both are modern, well designed, and genuinely competitive with the closed tools they’re up against. The only real caveat left comes down to data centers. Cloud AI has the ability to be much better than anything living on consumer hardware, and until my card magically turns into a rack of H100s, that’s where my subscription money keeps going.
